This article is part of our series that explore the business of artificial intelligence
Earlier this week, I wrote about the economy of GPT-3, OpenAI’s massive language model. I discussed what it would take for the company to make the AI algorithm profitable and secure future funding for its artificial general intelligence research.
Based on what we know, OpenAI faces a great challenge breaking even on the costs of developing, training, and running its huge neural network, let alone turning it into a revenue-generating business.
But now there’s a new twist to the story. On Tuesday, Microsoft declared it would be exclusively licensing GPT-3 as part of its ongoing partnership with OpenAI. This can give a big boost to the AI research lab’s search for a viable business model.
It can also break it.
GPT-3’s business model
To get a full account of the costs of developing and running GPT-3, read Monday’s post (and follow the discussion on Reddit for further information). But here’s a quick recap of the figures OpenAI faces:
- Above $10 million in expenses for research on GPT-3 and training the final model
- Tens of thousands of dollars in monthly cloud computing or server and electricity costs for running the model
- Possibly more than a million dollars in yearly retraining costs due to model decay
- Additional costs of customer support, marketing, IT, security, legal and other requirements of running a product. This could be in the tens of thousands of dollars based on the number and size of customers OpenAI acquires.
Note that OpenAI has released very little information about the process and these are estimates based on projections from the known costs of training and running other deep learning models on cloud and on-premise servers. OpenAI has a very close partnership with Microsoft and has helped develop a supercomputer for massive AI models. Microsoft has probably given the AI lab a much better deal to train and run GPT-3. Our equation also excludes the huge salaries OpenAI has to pay to its researchers for the sake of sticking to GPT-3 alone.
Based on the pricing plan declared by OpanAI, the company will need several dozen Build tier customers ($400/month) to break even on the costs of running GPT-3, and much more to cover the expenses of developing and training the model in a reasonable timeframe, and pay for the cost of retraining the language model every once in a while.
The problem is, GPT-3 is an entirely new type of technology, a language model that is capable of zero- and one-shot learning. There’s no precedent for it, and finding the right market for it is very difficult.
On the one hand, OpenAI will have to find areas where GPT-3 can create entirely new applications, such as content generation. On the other, it has to find markets where the language model can bring enough substantial improvement to existing applications to convince customers to switch from whatever technology they have to GPT-3.
According to OpenAI, GPT-3 has so far received tens of thousands of applications to date. But many of these people are interested in playing around with the free beta version. As soon as the company starts to charge for the GPT-3 application programming interface, then we’ll see how many of those applications have working business models. I don’t think much do.
Many products die because they can’t find the right product-market fit, the formula that allows them to acquire new customers in a cost-effective way. So, even if you have the most advanced and exciting technology in the world but don’t have access to the right market, you’re doomed to fail.
This is where Microsoft enters the picture.
Microsoft’s GPT-3 license
The Microsoft license was not a surprise. In 2019, when Microsoft invested $1 billion in OpenAI, the AI research lab declared (emphasis mine): “OpenAI is producing a sequence of increasingly powerful AI technologies, which requires a lot of capital for computational power. The most obvious way to cover costs is to build a product, but that would mean changing our focus. Instead, we intend to license some of our pre-AGI technologies, with Microsoft becoming our preferred partner for commercializing them.”
Here’s what Kevin Scott, Microsoft’s Chief Technology Officer, said on Tuesday: “Microsoft is teaming up with OpenAI to exclusively license GPT-3, allowing us to leverage its technical innovations to develop and deliver advanced AI solutions for our customers, as well as create new solutions that harness the amazing power of advanced natural language generation.”
I think the key here is the term “our customers.” Microsoft has a vast and established customer base. Over a million companies use Office 365, including 800,000 in English-speaking countries. Microsoft also owns Bing, the second-most-popular search engine, and the digital assistant Cortana. These are all areas that rely on natural language processing (NLP) technology.
There’s mutual benefit in the licensing. Through Microsoft, OpenAI will gain access to hundreds of millions of users and their interactions with the new AI technology. This will provide key insights and experience that will allow the research lab to improve GPT-3 and its successors.
But the real winner will be Microsoft, who will have exclusive access to the most advanced AI technology.
Per Scott: “The scope of commercial and creative potential that can be unlocked through the GPT-3 model is profound, with genuinely novel capabilities – most of which we haven’t even imagined yet. Directly aiding human creativity and ingenuity in areas like writing and composition, describing and summarizing large blocks of long-form data (including code), converting natural language to another language – the possibilities are limited only by the ideas and scenarios that we bring to the table.”
I can already imagine many use cases for GPT-3 in Word, Excel, OneNote, and Teams. But these are areas where Microsoft already dominates. I’m interested to see if Microsoft can use GPT-3 to move market dynamics in other areas where it currently lags.
In their book Prediction Machines, Ajay Agrawal, Joshua Gans, and Avi Goldfarb discuss how companies can leverage advances in machine learning can allow companies to restructure their operations in totally new ways, such as going from shopping-then-shipping to shipping-then-shopping.
An interesting area to observe is the search engine market. Bing currently has a little over 6 percent of the market share, far behind Google’s 87 percent. Will GPT-3 enable Microsoft to roll out new features that will redefine how we use search and draw more users to Bing? For instance, while you’re writing an article in Word, GPT-3-powered Bing could assist you with references to external resources. Basically, it could be a search engine that predicts and answers your questions in advance.
It would also be interesting to see how GPT-3 could change the voice assistant market. According to one study, Cortana has 22 percent of the market share, second behind Apple Siri’s 35 percent. Digital assistants are a bit trickier because they depend on devices and operating systems. Cortana is available on most operating systems and devices, but it is only the default option on Windows devices. Android devices come preinstalled Google assistant, and iOS devices with Siri. Amazon’s assistant is Alexa, but the company partnered with Microsoft to interoperate their assistants two years ago.
We have yet to see how Microsoft will expose users to improvements in Cortana, but the difference must be significant enough to convince them to give up their default assistants.
The key benefit of Microsoft’s GPT-3 license is that the company will not be bound to the API pricing model and will be subsidizing it for its customers. What this means is that Microsoft customers will probably get free access to GPT-3 features. Other companies will have to play with GPT-3 at OpenAI’s pricing model, which is at least 60 times higher than the cost of running the AI model.
The future of OpenAI
If Microsoft manages to get commercial benefits from GPT-3, it will probably provide the next round of funding to OpenAI. But there’s also a dark side to becoming too enmeshed with a large corporation. Microsoft’s first $1 billion came with the condition that part of it was in Azure credits, which means OpenAI could only spend it on Microsoft’s own cloud service. OpenAI also agreed to give Microsoft exclusive license to its advanced AI model.
What will be the new restrictions Microsoft sets on OpenAI’s research? Here are a few questions to ask:
- Will the exclusive license expand to future iterations of GPT and other technologies OpenAI develops?
- What happens to the user feedback data from GPT-3? If Microsoft holds exclusive ownership of the data, it can hold sway over OpenAI’s future research.
- Will Microsoft have any influence on the pricing model of the GPT-3 API? Microsoft might want to prevent OpenAI from lowering the price of the API to make it harder for other companies to create profitable products from the technology.
Overall, we can expect Microsoft to influence OpenAI to protect its own commercial interests. In return it will provide incentives, such as subsidized access to its cloud infrastructure and other benefits, such as helping with product management, cybersecurity, legal issues, and more challenges the research has not dealt with before. But this will make OpenAI even more dependent on Microsoft and submissive to its commercial goals. In the future, it can lead to OpenAI’s acquisition by Microsoft.
As I argued last month, the partnership has already resulted in the transformation of OpenAI from a scientific research lab to a hybrid between its former self and a commercial company. It is now walking a fine line between sticking to its scientific, nonprofit goal of developing human-level AI and keeping its quasi-owner Microsoft happy.